Signed-off-by: Onur Yilmaz <oyilmaz@nvidia.com>

Bug fixes:
1. conv_state save-before-read: extract initial conv states BEFORE
   causal_conv1d_varlen_states + tensor_masked_update overwrites the
   conv_state buffer. Previously, initial_conv_states was read AFTER
   the buffer was updated, so restored requests would see their own
   newly-computed final states instead of the pre-existing initial
   states, corrupting the convolution output.

2. cu_chunk_seqlens OOB: the SSM Triton kernels allocate per-chunk
   output arrays of size chunk_size (128). Passing cu_seqlens directly
   as cu_chunk_seqlens caused out-of-bounds memory access when any
   sequence exceeded chunk_size tokens. Fix: subdivide each sequence
   into chunks of at most self.chunk_size, producing correct
   cu_chunk_seqlens boundaries.

3. zxBCdt padding mismatch: after conv1d, the per-request loop rebuilt
   xBC with only real tokens while dt and z retained padded token count.
   This caused a shape assertion failure in the SSM kernel. Fix: strip
   padded tokens from zxBCdt before _ssm_prefill, then pad the output
   back to the original padded size for downstream residual add.

4. Per-request conv1d with initial_states: causal_conv1d_fn cannot
   accept both seq_idx and initial_states simultaneously. The old code
   passed seq_idx to handle multiple sequences but this zeroes state at
   sequence boundaries instead of using the cached initial states. Fix:
   loop over requests, calling causal_conv1d_fn per-request with
   initial_states and channels-last layout.

Improvements:
- Unify all Mamba prefill (including chunked) through single varlen SSM
  kernel call, removing separate chunked-prefill routing and the
  _batch_indices_chunked_prefill / _device_chunked_prefill metadata
- Simplify _dynamic_inference to flat decode + prefill structure
- Add _dynamic_inference_prefill helper that strips CUDA-graph padding
  from metadata and data tensors before calling _ssm_prefill
- Remove deprecated constructor parameters (use_mem_eff_path, d_state,
  headdim, ngroups) and their warnings
- Add assertion format string in ssd_combined.py for easier debugging

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Instead of breaking prefill into chunks at Mamba state boundaries,
process full prefill in one kernel call and extract SSM/conv states
at specified token offsets. This eliminates the dependency on chunked
prefill scheduling and simplifies the engine.

Key changes:
- Two-map hash design: kv_hash_to_block_id + mamba_hash_to_block_id
- Mamba cache infrastructure: GPU memory pool for SSM/conv states
- Coupled prefix matching: skip tokens limited by Mamba match count
- Intermediate offset computation at KV divergence and last-aligned
- Engine passes mamba match count, commits states after forward pass
- KV eviction automatically invalidates Mamba state
- 18 new tests covering all Mamba caching paths

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>